It might be important that "impossible" depends on your state of knowledge. For example in counterfactual mugging getting the large reward doesn't become impossible until after you know the result of the coin flip. So you should perhaps more carefully state your principal as "impossible at the time that the agent make the decision".

Is there an argument why IIO should overrule reflective consistency in Counterfactual Mugging?

In this post http://lesswrong.com/lw/f37/naive_tdt_bayes_nets_and_counterfactual_mugging/ I argued, as an aside, for another principle: that it shouldn't matter whether you were being simulated or whether anyone was simply predicting the result of that simulation.

Then (very roughly) CDT behaves as EDT/UDT if you assume that Newcomb is simulating you, because you can't tell whether you're the simulation or the "real" you. But this argument also argues for the counterfactual mugging, unlike yours.